Google opened its Gemini 4 Argon model to developers and businesses worldwide on 13 October 2026. The price shown at launch, 2 dollars per million input tokens and 10 dollars per million output tokens, is an introductory price. A note at the bottom of Google's announcement sets its end date as 30 November 2026: from 1 December, the same usage will be billed at 4 and 20 dollars. A team that costs a project based on October prices therefore underestimates the per-token cost by half from 1 December.

Who has had access to Argon since 13 October

Google opened Argon on 13 October to developers and businesses via the API, AI Studio and its Cloud agent platform. The model is also arriving in Gemini for business, in Antigravity for Ultra subscribers, and gradually in the Gemini app for Pro and Ultra subscribers.

Access in the Gemini app and integration into software are separate offerings. The API pricing table reserves Argon for the paid tier. According to the subscription documentation, the Pro plan does not include it in AI Studio, while Ultra allows prototyping there according to its quotas. To cost a production application, you need to look at the terms of the API project used.

A price schedule that doubles on all token lines

The pricing table consulted on 16 October distinguishes introductory prices from those applicable from 1 December. All amounts below are in dollars per million tokens:

  • standard: 2 dollars input and 10 dollars output until 30 November, then 4 and 20 dollars;
  • input tokens read from cache: 0.10 dollar, then 0.20 dollar;
  • batch processing and Flex mode: 1 and 5 dollars, then 2 and 10 dollars;
  • Priority mode: 3.60 and 18 dollars, then 7.20 and 36 dollars.

For a fixed volume of 10 million non-cached input tokens and one million output tokens, standard mode thus costs 30 dollars during the promotion, then 60 dollars. This example excludes any tool fees and assumes the same volume of reasoning and responses.

The output price includes reasoning tokens produced before the answer. Their volume depends in particular on the reasoning level chosen: low, medium or high. The doubling of the unit price is certain; how the bill evolves will also depend on the number of tokens consumed.

On the cache side, according to the Google Cloud developer guide, Argon only supports implicit caching, applied automatically to common prompt prefixes of at least 6,144 tokens. Explicit caching, which allows you to reserve a context and control its reuse, is not supported.

The EU endpoint adds 10% to the global rate

On Google Cloud, the model guide offers global, US and EU endpoints. The Cloud pricing table applies a 10% surcharge outside global. For EU, the promotional cost is equivalent to 2.20 dollars input and 11 dollars output, then 4.40 and 22 dollars from 1 December. This price difference alone does not describe the contractual guarantees on data residency.

The discount mechanism also differs by channel. On Google Cloud, the pricing note provides credits equal to 50% of net spend during the promotional period. A buyer must therefore check the amount billed, the credits granted and how they appear in exports, rather than assuming the displayed unit price will be directly halved.

Price per token is not enough to compare models

As of 16 October, Argon's introductory price equals that of Claude Sonnet 5.5 and GPT-6.1 Sol in short context, for non-cached input and standard-mode output. The price announced for December corresponds to the prices currently charged for Claude Opus 5.5, i.e. 4 and 20 dollars. Cache options, tools and context volumes can alter this comparison.

The Vals Index illustrates the gap between unit price and cost per task. In its update of 13 October, it calculates for Argon an average cost per test of 15.68 dollars at the full price of 4 and 20 dollars, compared with 21.34 dollars for Sonnet 5.5 and 32.14 dollars for Opus 5.5. Argon consumes fewer tokens on this benchmark. These amounts depend on the tasks and settings of Vals' protocol; they do not predict the bill for a given application.

Points to settle before 1 December

For a team evaluating Argon, the introductory window is mainly for measurement. The real cost of a use case depends on the number of reasoning tokens consumed at the chosen reasoning level, the hit rate of implicit caching and the region of the call. These three parameters can be measured on test calls during the promotional period and used to project the bill at December prices.

Our articles will then appear first in Google Top Stories.